By Ryan Mullins, Research Engineer and RAI Toolkit Tech Lead OCT 23, 2024
Building AI responsibly is crucial. That’s why we created the Responsible GenAI Toolkit, providing resources to design, build, and evaluate open AI models. And we’re not stopping there! We’re now expanding the toolkit with new features designed to work with any LLMs, whether it’s Gemma, Gemini, or any other model. This set of tools and features empower everyone to build AI responsibly, regardless of the model they choose.
Here’s what’s new:
SynthID Text: watermarking and detecting AI-generated content
Is it difficult to tell if a text was written by a human or generated by AI? SynthID Text has you covered. This technology allows you to watermark and detect text generated by your GenAI product.
How it works: SynthID watermarks and identifies AI-generated content by embedding digital watermarks directly into AI-generated text.
Open source for developers: SynthID for text is accessible to all developers through Hugging Face and the Responsible GenAI Toolkit.
Learn more:
- Dive into the full technical details in the Nature paper.
- Discover how to apply SynthID responsibly on the Responsible GenAI Toolkit website.
- Explore the SynthID Technology Page for a comprehensive overview of SynthID’s capabilities across all modalities.
Use it today:
- A production-grade implementation is available in the Hugging Face Transformers library. Check out the Hugging Face Space for a step-by-step guide to configuring SynthID for your GenAI applications.
- A reference implementation is also available on GitHub and PyPI, with a Colab Notebook for interactive learning.
We invite the open source community to help us expand the reach of SynthID Text across frameworks, based on the implementations above. Reach out on GitHub or Discord with questions.
Model Alignment: refine your prompts with LLM assistance
Crafting prompts that effectively enforce your business policies is crucial for generating high-quality outputs.
The Model Alignment library helps you refine your prompts with support from LLMs.
Provide feedback about how you want your model’s outputs to change as a holistic critique or a set of guidelines.
Use Gemini or your preferred LLM to transform your feedback into a prompt that aligns your model’s behavior with your application’s needs and content policies.
Use it today:
- Experiment with the interactive demo in Colab and see how Gemini can help align and improve prompts for Gemma.
- Access the library on PyPI.
- Utilize these methods in Vertex AI Studio with Refine Prompt.
Prompt Debugging: streamline LIT deployment on Google Cloud
Debugging prompts is essential for responsible AI development. We’re making it easier and faster with an improved deployment experience for the Learning Interpretability Tool (LIT) on Google Cloud.
- Efficient, versatile model serving: Leverage LIT’s new model server container to deploy any Hugging Face or Keras LLM with support for generation, tokenization, and salience scoring on Cloud Run GPUs.
- Expanded connectivity from the LIT App: Seamlessly connect to self-hosted models or Gemini via the Vertex API (generation only).
- Learn more: Explore LIT’s capabilities for responsible model alignment on the Responsible GenAI Toolkit website.
Info from Google for Developers
Ryan Mullins
I am a Software Engineer on the People + AI Research team, part of Google Research, where I study human-AI collaboration and AI/ML interpretability and explainability. Previously I was a Senior Researcher at Aptima, where I studied human-AI collaboration for software security, planning, and information analysis.
I research and develop visualizations, experiences, and analytics for Human-AI systems working across a variety of domains, including cybersecurity, planning, and problem-solving.
